CN107341768A - Grid noise reduction - Google Patents
Grid noise reduction Download PDFInfo
- Publication number
- CN107341768A CN107341768A CN201610285481.4A CN201610285481A CN107341768A CN 107341768 A CN107341768 A CN 107341768A CN 201610285481 A CN201610285481 A CN 201610285481A CN 107341768 A CN107341768 A CN 107341768A
- Authority
- CN
- China
- Prior art keywords
- normal direction
- dough sheet
- geometric properties
- true
- grid
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000009467 reduction Effects 0.000 title claims abstract description 32
- 238000013507 mapping Methods 0.000 claims abstract description 44
- 238000000034 method Methods 0.000 claims description 75
- 238000012549 training Methods 0.000 claims description 62
- 239000013598 vector Substances 0.000 claims description 28
- 230000008859 change Effects 0.000 claims description 20
- 238000000611 regression analysis Methods 0.000 claims description 16
- 238000001914 filtration Methods 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 14
- 230000002146 bilateral effect Effects 0.000 claims description 6
- 230000004044 response Effects 0.000 claims description 5
- 230000000149 penetrating effect Effects 0.000 claims 1
- 230000006870 function Effects 0.000 description 13
- 238000004891 communication Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 8
- 230000009471 action Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 239000004576 sand Substances 0.000 description 4
- 238000005070 sampling Methods 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000006855 networking Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 108010022579 ATP dependent 26S protease Proteins 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/20—Finite element generation, e.g. wire-frame surface description, tesselation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/20—Finite element generation, e.g. wire-frame surface description, tesselation
- G06T17/205—Re-meshing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/20—Image enhancement or restoration using local operators
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/60—Image enhancement or restoration using machine learning, e.g. neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2200/00—Indexing scheme for image data processing or generation, in general
- G06T2200/04—Indexing scheme for image data processing or generation, in general involving 3D image data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20024—Filtering details
- G06T2207/20028—Bilateral filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Graphics (AREA)
- Medical Informatics (AREA)
- Computational Linguistics (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Algebra (AREA)
- Probability & Statistics with Applications (AREA)
- Image Processing (AREA)
- Complex Calculations (AREA)
- Image Analysis (AREA)
Abstract
In the disclosure, it is proposed that a kind of scheme being used for surface mesh noise reduction.For the surface mesh including polygonal patch, the noise normal direction of the first dough sheet in grid and true normal direction are obtained.Then, based on noise normal direction, the first geometric properties of the first dough sheet are determined from multiple neighbouring dough sheets of the first dough sheet in grid.Then, based on the first geometric properties and true normal direction, the mapping of the true normal direction of the first geometric properties to the first dough sheet is determined, for grid noise reduction.
Description
Background technology
The appearance of three-dimensional (3D) scanner and depth camera greatly simplifies Geometric Modeling process,
And allow one to easily obtain 3D shapes from real world.With it is artificial by artist
Unlike the 3D grids (mesh) of establishment, the noise in scanned 3D grids is usual
Than more significant." noise " mentioned here refers to the process of 3D grids in scanned or intake
In suffered random signal interference.The reason for causing these noises be it is many, it is all if any
Error in the equipment precision of limit, the material properties of real world objects, process of reconstruction, etc..
In fact, in 3D modeling and each stage of processing, all there may be reduces grid noise
Demand.Therefore, noise is removed from scanned grid turns into important in 3D geometric manipulations
Task.
The content of the invention
According to the realization of subject matter described herein, it is proposed that a kind of grid drop based on data-driven
The scheme made an uproar.According to the program, for the grid including dough sheet, the first face in grid is obtained
The noise normal direction and true normal direction of piece.Then, based on noise normal direction, from the first face in grid
Multiple neighbouring dough sheets of piece determine the first geometric properties of the first dough sheet.Then, based on more than the first
What feature and true normal direction, determines the mapping of the true normal direction of the first geometric properties to the first dough sheet,
For to grid noise reduction.
There is provided Summary be in order to introduce the selection to concept in simplified form, it
Will be further described in specific implementation below.Summary is not intended to identify and wanted
The key feature or principal character of the theme of protection are sought, is also not intended to limit theme claimed
Scope.
Brief description of the drawings
Fig. 1 shows the environment 100 for the realization that can be implemented within subject matter described herein
Block diagram;
Fig. 2 shows the method being used for grid noise reduction of the realization according to subject matter described herein
200 flow chart;
Fig. 3 shows the method for generating regression tree of the realization according to subject matter described herein
300 flow chart;And
Fig. 4 shows the regression tree for being used to generate cascade of the realization according to subject matter described herein
Method 400 flow chart;
Fig. 5 shows the method being used for grid noise reduction of the realization according to subject matter described herein
500 flow chart;
Fig. 6 shows the device being used for grid noise reduction of the realization according to subject matter described herein
600 block diagram;And
Fig. 7 is shown in which that showing for one or more realizations of subject matter described herein can be implemented
The block diagram of example computing system/server 700.
In the accompanying drawings, same or analogous reference symbol is used to indicate that same or analogous element.
Specific embodiment
Subject matter described herein is discussed now with reference to some example implementations.It should be appreciated that by
Stated these realize merely to so that those of ordinary skill in the art better understood when and
Therefore subject matter described herein is realized, rather than implies any restrictions to the scope of this theme.
As used herein, term " comprising " and its variant will be read as meaning " bag
Include but be not limited to " open-ended term.Term "based" will be read as " at least in part
Based on ".Term " realization " and " one kind is realized " will be read as " at least one reality
It is existing ".Term " another realization " will be read as " at least one other realization ".Term
" first ", " second " etc. may refer to different or identical object.Further below may
Including other clear and definite and implicit definition.Unless additionally explicitly pointing out, the definition of term is passed through
It is consistent to wear specification.
In addition, as used herein, term " grid (mesh) " refers to 3D geometry moulds
A kind of representation of type.Point sampling can be carried out to the object in 3D coordinate systems, this
A little sampled points can be connected to a series of facet in certain sequence, and (triangle is coplanar
Quadrangle, pentagon etc.), these facets are also referred to as dough sheet (facet).So, pass through
The object of sampling can be made up of the dough sheet of many, as net, therefore be referred to as grid.
In the context of the disclosure, the realization of subject matter described herein will be illustrated mainly in combination with grid
Principle and thought.In addition to grid, 3D geometrical models can also use point cloud (point cloud)
To represent.Point cloud is the geometrical model using discrete sampling point as primitive.It can utilize any current
A cloud is redeveloped into grid by known or exploitation in the future method.For example, it can utilize in point cloud
Some neighbouring sampled points construct facet (be similar to dough sheet in grid).Therefore, originally
The realization of theme described by text will also apply to a cloud.
As described above, it is the weight in geometric manipulations that noise is removed from scanned grid
Want task.The basic goal of grid noise reduction is to recover it from the noisy surface of band of object
The truly surface of (ground-truth).Most of existing noise reduction algorithm is based on wave filter
Method.For example, the specific wave filter of such as two-sided filter can be designed, and apply trial and error
Method (trial and error) selects the parameter of wave filter to reach gratifying noise reduction.
However, this whole process is cumbersome.In addition, existing noise reduction algorithm usually assumes that noise
Statistical nature.For example, it is assumed that noise is class Gaussian Profile (Gaussian-like), this is true
It is not necessarily in real data correctly.For from many factors (such as scanning device, scanning
Environment, object materials etc.) noise for, noise be characterized in complexity.Thus, it is supposed that
The existing noise reduction algorithm of the statistical nature of noise can not effectively remove these noises.
In order to solve the above problems and other potential problems in one or more, according to this
A kind of example implementation of theme described by text, it is proposed that scheme being used for grid noise reduction.Such as with
Described by upper, grid can be made up of many dough sheets, and each dough sheet can have perpendicular to this
The straight line (also referred to as " normal ") of dough sheet, the direction of the straight line can be referred to as " normal direction
(normal)”.Therefore, the process of grid noise reduction can be included a series of faces in network
The noisy normal direction of band (hereinafter also referred to " noise normal direction ") of piece reverts to corresponding true
Normal direction (ground-truth normal), and update the top of grid based on these true normal direction
The process of point position.
Fig. 1 shows the environment 100 for the realization that can be implemented within subject matter described herein
Block diagram.It is to be understood that the 26S Proteasome Structure and Function of environment 100 is described merely for exemplary purpose
Rather than imply any restrictions of the scope for subject matter described herein.Subject matter described herein
It can be embodied in different structure and/or function.
As shown in figure 1, environment 100 can generally include model training systems 110 and model
Application system 120., can be with to the scheme of grid noise reduction according to the realization of subject matter described herein
Including two stages:Model training stage and model application stage.
In the model training stage, model training systems 110 can be configured as receiving training number
According to 130, and based on training data 130 come generation model 140.Training data 130 can be with
The number being made up of the noise normal direction and corresponding true normal direction of a series of dough sheets in grid
According to set.Can by model training systems 110 based on training data 130 come generate be used for pair
The model 140 of grid noise reduction.Model 140 can be represented from the noise normal direction of dough sheet to true method
To mapping.
In the model application stage, model application system 120 can be configured as receiving model 140
With input data 150, and output data is generated based on model 140 and input data 150
160.Input data 150 can represent one group of noise normal direction of grid.Model application system 120
It can utilize model 140 that this group of noise normal direction is mapped into corresponding one group of true normal direction, i.e.,
Output data 160.Based on output data 160, the vertex position of grid can be updated, so as to
Realize the noise reduction to grid.
Fig. 2 shows the method being used for grid noise reduction of the realization according to subject matter described herein
200 flow chart.For example, method 200 can be by model training systems 110 as shown in Figure 1
To perform.It should be appreciated that method 200 can also include unshowned additional step and/
Or shown step can be omitted.The scope of subject matter described herein is not limited in this respect.
Method 200 starts from step 210.In step 210, by model training systems 110
Obtain the noise normal direction of one or more of grid dough sheet and true normal direction.For convenience
For the sake of, the dough sheet handled herein is referred to as " the first dough sheet ".The noise normal direction of first dough sheet and
True normal direction may be constructed training data 130 as shown in Figure 1.For example, user can be based on
The mode of input obtains training data 130, can also be obtained by way of Data Synthesis
Training data 130.Further, it is also possible to obtained from any equipment for being currently known or developing in the future
Obtain training data 130.For example, the first face can be obtained using the scanning device of lower accuracy
The noise normal direction of piece and obtain the true of the first dough sheet using the scanning device of degree of precision
Normal direction.The scope of subject matter described herein is not limited in this respect.
Next, method 200 proceeds to step 220.In a step 220, by model training
System 110 is based on noise normal direction, and is determined from multiple neighbouring dough sheets of the first dough sheet in grid
First geometric properties of one dough sheet.In some implementations, " neighbouring dough sheet " can include and the
One dough sheet shares the dough sheet on summit and/or side.In some implementations, " neighbouring dough sheet " can wrap
The distance between its geometric center (centroid) and the geometric center of the first dough sheet is included to be less than in advance
Determine the dough sheet of threshold value.In addition, in some implementations, " neighbouring dough sheet " can include in grid
All dough sheets.In some implementations, can be between the first dough sheet and multiple neighbouring dough sheets
Normal direction changes and change in location is filtered, and the characteristic vector of the first geometric properties is represented with generation.
For example, multiple neighbouring dough sheets can be selected based on the coverage of filtering parameter.
By taking 3D triangle griddings (that is, each dough sheet in grid is tri patch) as an example, net
Lattice can be represented as M=(V, F), wherein V={ viRepresent summit set and
F={ fiRepresent dough sheet set.Dough sheet fiGeometric center can be represented as ci, and face
Piece fiNormal direction can be represented as ni。
In some implementations, the normal direction between the first dough sheet and multiple neighbouring dough sheets can be changed
Bilateral filtering is carried out with change in location to generate the characteristic vector of the first geometric properties of expression.For example,
(k+1) secondary iteration of two-sided filter (bilateral filter) can be defined as foloows:
Wherein njRepresent dough sheet fjInitial noisc normal direction.Represent the dough sheet f by k iterationiWarp
The normal direction of filtering, when k is 0,Represent dough sheet fiInitial noisc normal direction.AjExpression face
Piece fjArea.In formula (1) (and following all formula), operator " " represents
Product.WsAnd WrIt is to characterize the position similitude and normal direction similitude between a pair of dough sheets respectively
The weighting function of monotone decreasing.It is, for example, possible to use Gaussian function Wσ(x)=exp (- x2/
(2σ2)), then WsAnd WrStandard deviation be respectively σsAnd σr。 It is normalization item.When suitably setting σs、σrDuring with iteration number,
The noise in normal direction can effectively be filtered while the sharp features of grid are retained.Can
So that formula (1) to be expressed asTo bind parameter σsAnd σr, it is hereby achieved that dough sheet
fiOn characteristic vector Si.That is, SiIt can be represented using formula below (2):
WhereinIt is the parameter set derived comprising various Gausses, l is represented
Parameter included in parameter set P is to (σs,σr) number.Using formula (2) come the S that representsi
Normal direction descriptor (the bilateral filtered normal through bidirectional filtering can be referred to as
descriptor,B-FND)。
Alternatively or additionally, in some implementations, can be with order to tackle higher noise level
Filter is guided to the normal direction change between the first dough sheet and multiple neighbouring dough sheets and change in location
Ripple represents the characteristic vector of the first geometric properties to generate.Wave filter (guided filter)
Also referred to as combine two-sided filter (joint bilateral filter).Due to wave filter to
Reliable guiding (guidance) is introduced in two-sided filter, therefore for high-level noise more Shandong
Rod.For example, wave filter can be defined as foloows:
Wherein g () represents the guiding of predefined dough sheet normal direction.For example, in one embodiment, can
To use Gaussian processes to wave filter as guiding, i.e., WhereinFor
Normalize item.
Formula (3) can be expressed asIt is hereby achieved that dough sheet fiOn feature
Vectorial Si.That is, SiIt can be represented using formula below (4):
Using formula (4) come the S that representsiThe normal direction descriptor of channeled filtering can also be referred to as
(guided filtered normal descriptor,G-FND).For the net with larger error
For lattice, G-FND in principal character is distinguished than B-FND more robusts, and B-FND for
Weaker aspect ratio G-FND is more sensitive.
It should be appreciated that in addition to above-mentioned two-sided filter and wave filter, can be with profit
With the wave filter of other robusts the characteristic vectors of the first geometric properties is represented to generate.Retouched herein
The scope for stating theme is not limited in this respect.
By collecting the normal direction descriptor of all dough sheets in grid, set of eigenvectors can be obtained
Close (being expressed as Ω).Characteristic vector S in ΩiInclude a series of filtered dough sheet normal direction.
Additionally or alternatively, the characteristic vector in Ω can be further processed.For example,
Can be to the single feature vector S in ΩiAlignd.Assuming that SiInclude d filtered methods
To:m1,m2,…,md.Normal direction tensor can be builtMatrix TiThree
Characteristic vector is defined from SiIn filtered dough sheet normal direction residing for local coordinate system to the world
The spin matrix R of coordinate system (that is, the coordinate system residing for grid)i.By by each mjIt is multiplied byCharacteristic vector SiIt can be aligned.After all characteristic vectors in Ω are aligned,
Principal component analysis (Principle Component Analysis, PCA) can be performed to reduce spy
Levy the dimension of vector.The dimension of characteristic vector is used for the number for referring to the item included in characteristic vector
Mesh (for example, d).The number of principal component can by keeping characteristics energy at least 90% come really
It is fixed, for example, can be determined that 3.Further, it is also possible to using clustering algorithm to through dimensionality reduction or
The characteristic vector of non-dimensionality reduction is clustered.
Next, method 200 proceeds to step 230.In step 230, by model training
System 110 is based on the first geometric properties and true normal direction, determines the first geometric properties to the first face
The mapping of the true normal direction of piece, for grid noise reduction.
In some implementations, can be determined using regression analysis (regression analysis)
The mapping of the true normal direction of first geometric properties to the first dough sheet.Regression analysis is to be used to determine two
A kind of method of dependence between kind or two or more variables.For example, regression analysis can lead to
The causality crossed between the geometric properties of regulation dough sheet and true normal direction establishes regression model,
And according to training data come the parameters of solving model, then evaluate regression model and whether can
It is enough to be fitted training data well.
As an example, in some implementations, can be represented by generating the regression tree of the mapping come
Determine the mapping of the true normal direction of the first geometric properties to the first dough sheet.In this regard, Fig. 3 shows
The flow for being used to generate the method 300 of regression tree of the realization according to subject matter described herein is gone out
Figure.Method 300 can be considered as a kind of realization of the step 230 in method 200, and for example
It can be performed by model training systems 110 as shown in Figure 1.
In step 310, at each node of regression tree, the instruction associated with the node is utilized
Practice data set to train the first regression model.For example, at the root node of regression tree, number is trained
The first geometric properties of each dough sheet in multiple dough sheets in grid and true can be included according to collection
Real normal direction.Training dataset at root node can be represented asWherein SiTable
Show dough sheet fiCharacteristic vector,Represent dough sheet fiTrue normal direction,Represent training data
One sample of collection, and N represents the number for the sample that training data is concentrated.
For example, in some embodiments it is possible to select Single hidden layer feedforward neural networks
(Single-hidden Layer Feed Forward network, SLFN) model is used as recurrence
Model, and extreme learning machine (extreme learning machine, ELM) skill can be based on
Art trains the regression model.ELM essence is that the weight of hidden layer can be given birth at random
Into and without adjustment.For the training dataset with N number of sample, there are L to be hidden section
The ELM of point can be represented using formula below (5):
Wherein i=1 ..., N.Dough sheet normal direction desired from recurrence is represented, it is one three
Dimensional vector.Export weight vectorsIn each weight withA dimension be associated.θ(aj,bj,Si) represent concealed nodes activation primitive.For example,
θ(aj,bj,Si) S type functions (sigmoid function) are can be configured to, and can utilize
Formula below (6) represents:
Wherein ajAnd bjFor respectively from interval [- 1,1]M[0,1] randomly selected concealed nodes weight, M
It is vectorial SiDimension.Export weight vectorsIt is unknown in ELM, it can be by asking
Solve LpNorm regularization least square problem determines.Can using formula below (7) come
Represent:
The cost function returned can be represented using formula below (8):
For example, the norm p in formula (7) and formula (8) can be arranged to 1, so as to pass through
WillIt is optimized for L1What the error that norm defines was concentrated to reduce training data
Negatively affected caused by outlier.Iteration weight weighted least-squares method can be employed to find
The minimum value of weight is exported, it can be represented as formula below (9):
Wherein l represents iterations.Once output weight, it is known that if ELM can predict removal noise
Dough sheet normal direction.
In some embodiments it is possible at each node of regression tree, based on ELM technologies
To train the first regression model, the error of fitting of the first regression model for example can be according to above-mentioned public affairs
Formula (8) is calculated as E0。
In step 320, multiple training data subsets are obtained from training dataset.For example, as above
It is described, training dataset can be divided into multiple clusters using PCA.Certainly, any other
Appropriate technology or algorithm can be used in combination with the embodiment of theme described here, for inciting somebody to action
Training dataset is divided into multiple clusters.For example, the clustering method for being currently known or developing in the future
It can be used.
In step 330, multiple second regression models are trained using multiple training data subsets.
For example, multiple second can be trained based on nerual network technique for multiple training data subsets
Regression model.Train the mode of the second regression model and the above described instruction in the step 310
It is identical to practice the mode of the first regression model, is no longer described in further detail herein.
Next, method 300 is carried out to step 340.In step 340, in response to multiple
The error of fitting sum of two regression models is less than predictive error threshold value, generates more height sections of node
Point.For example, if the error of fitting sum of multiple second regression models is less than E02%, then
Multiple child nodes of node can be generated.Each child node is associated with a training data subset.
Next, in step 350, determine the leaf node of regression tree (that is, without sub- section
The node of point) whether number reach predetermined threshold.For example, predetermined threshold can be configured to 8.
When the leaf segment of regression tree is counted out more than or equal to 8, the regression tree stops growing (that is, square
Method 300 terminates);When the leaf segment of regression tree is counted out less than 8, method 300 is iteratively performed
Step 310 to 350, counted out until the leaf segment of regression tree and reach 8.
It should be appreciated that in addition to above-mentioned SLFN models, other models can also be utilized to make
For regression model.The scope of subject matter described herein is not limited in this respect.
In some implementations, the first geometric properties can be determined by generating the regression tree of cascade
To the mapping of the true normal direction of the first dough sheet.By the cascade of multiple regression trees, can realize more
Small error of fitting (that is, lower cost functional value).In this regard, Fig. 4 shows root
According to the flow for being used to generate the method 400 of the regression tree of cascade of the realization of subject matter described herein
Figure.Method 400 can be considered as a kind of realization of the step 230 in method 200, and for example
It can be performed by model training systems 110 as shown in Figure 1.
In step 410, based on the first geometric properties and true normal direction, generation represents the of mapping
First geometric properties are mapped to the true method for representing the first dough sheet by one regression tree, the first regression tree
To the first normal direction.For example, the can be generated using with reference to Fig. 3 come the method 300 described
One regression tree.
In step 420, based on the first normal direction, the second geometric properties of the first dough sheet are determined.Example
Such as, can utilize with reference to Fig. 2 come the step 220 identical mode of the method 200 described come
It is determined that represent the characteristic vector of the second geometric properties of the first dough sheet.
Next, method 400 is carried out to step 430.In step 430, based on the second geometry
Feature and true normal direction, the second regression tree of generation and the cascade of the first regression tree.For example, can be with
Using Fig. 3 is combined the second regression tree is generated come the method 300 described.For example, second returns
Second geometric properties can be mapped to the second normal direction of the true normal direction for representing the first dough sheet by tree,
And the error between the second normal direction and true normal direction is less than between the first normal direction and true normal direction
Error.
It should be appreciated that in order to further reduce error of fitting, can also further generate and second
3rd regression tree of regression tree cascade, the 4th regression tree cascaded with the 3rd regression tree etc..Institute
Generation the regression tree through cascade number for example can by the threshold value of predetermined error of fitting Lai
Determine, the scope of subject matter described herein is not limited in this respect.In addition, in order to handle tool
There is the training dataset of larger error, it is, for example, possible to use G-FND trains the first recurrence
Tree, to cause the output of the first regression tree that there is smaller error.It is then possible to based on first time
B-FND is extracted in Gui Shu output.Because B-FND is more sensitive for the sharp features of grid,
Therefore the second regression tree cascaded with the first regression tree can be trained using B-FND.
Therefore, by above in association with the method 200,300 and/or 400 described by Fig. 2 to Fig. 4,
In the model training stage, model training systems 110 have been based on training data 130 and generated
Model 140 (for example, by single regression tree or multiple recurrence tree representations of cascade).In mould
In the type application stage, model application system 120 can utilize generated model 140 by table
Show that the input data 150 of one group of noise normal direction of grid is mapped to one group of true method corresponding to expression
To output data 160.
Fig. 5 shows the method being used for grid noise reduction of the realization according to subject matter described herein
500 flow chart.For example, method 500 can be by model application system 120 as shown in Figure 1
To perform.It should be appreciated that method 500 can also include unshowned additional step and/
Or shown step can be omitted.The scope of subject matter described herein is not limited in this respect.
In step 510, one or more of grid dough sheet is obtained by model application system 120
Noise normal direction.For the sake of for convenience, the dough sheet handled herein is referred to as " the second dough sheet ".
The noise normal direction of second dough sheet can form input data 150 as shown in Figure 1.
Next, method 500 is carried out to step 520.In step 520, based on the second dough sheet
Noise normal direction, determine the of the second dough sheet from multiple neighbouring dough sheets of the second dough sheet in grid
Three geometric properties.For example, can utilize with reference to Fig. 2 come the step 220 of method 200 that describes
Identical mode come determine represent the second dough sheet the 3rd geometric properties characteristic vector, herein not
It is described in further detail again.
Then, in step 530, based on the 3rd geometric properties and mapping, the second dough sheet is determined
True normal direction.The true normal direction of second dough sheet can form output data 160 as shown in Figure 1.
As described above, for example, the mapping of the geometric properties of dough sheet to true normal direction can be by single recurrence
Set to represent.In this example, intermediate node that can be based on regression tree (that is, has son section
The node of point) determine to represent that the characteristic vector of the 3rd geometric properties of the second dough sheet is branched
Path.When the characteristic vector for the 3rd geometric properties for representing the second dough sheet reaches the leaf of the regression tree
, can be based on the regression model (that is, formula (5)) corresponding to leaf node come really during node
The true normal direction of fixed second dough sheet.
In some implementations, the mapping of the geometric properties of dough sheet to true normal direction can be by cascading
Multiple regression trees represent.For example, it is assumed that the geometric properties of dough sheet to true normal direction mapping by
The first regression tree and second regression tree of cascade represent.In this example, can be as described above
To perform method 500, determined with the noise normal direction based on the second dough sheet and the first regression tree
3rd normal direction of the second dough sheet.It is then possible to method 500 is repeatedly carried out, with based on second
The 3rd normal direction and the second regression tree of dough sheet determines the true normal direction of the second dough sheet.For example, can
To obtain the 3rd normal direction (510) of the second dough sheet.It is then possible to based on the second dough sheet
Three normal direction determine the 4th geometric properties (520) of the second dough sheet.Finally, the 4th can be based on
Geometric properties and the second regression tree determine the true normal direction (530) of the second dough sheet.
Additionally or alternatively, can be with after the true normal direction of the dough sheet in obtaining grid
The vertex position of grid is updated using these true normal direction.For example, can by solve such as with
Under the non-linear minimisation problem that represents of formula (10) obtain updated vertex position:
Wherein V represents the set on summit,Represent the true normal direction of the dough sheet in grid, ejkRepresent
Dough sheet fiWith end points vjAnd vkSide, and Ai(V) dough sheet f is representediArea.
Fig. 6 shows the device being used for grid noise reduction of the realization according to subject matter described herein
600 block diagram.Model training systems 110, model application system 120 or both can be by
Device 600 is realized.As shown in fig. 6, device 600 can include training data acquisition module 610,
It is configured as obtaining the noise normal direction of the first dough sheet in grid and true normal direction.Device 600 is also
The first geometric properties determining module 620 can be included, be configured as being based on noise normal direction, from net
Multiple neighbouring dough sheets of the first dough sheet in lattice determine the first geometric properties of the first dough sheet.In addition,
Device 600 can also include mapping determining module 630, be configured as being based on the first geometric properties
With true normal direction, determine the mapping of the true normal direction of the first geometric properties to the first dough sheet, with
In to grid noise reduction.
Additionally or alternatively, device 600 can also include input data acquisition module, quilt
It is configured to obtain the noise normal direction of the second dough sheet in grid.Device 600 can also include the 3rd
Geometric properties determining module, the noise normal direction based on the second dough sheet is configured as, from grid
Multiple neighbouring dough sheets of second dough sheet determine the 3rd geometric properties of the second dough sheet.In addition, device
600 can also include output data determining module, are configured as based on the 3rd geometric properties and reflect
Penetrate, determine the true normal direction of the second dough sheet.
For purposes of clarity, some optional modules of device 600 are not shown in figure 6.
It will be appreciated, however, that it is equally applicable to device above with reference to each feature described by Fig. 1-5
600.Moreover, the modules of device 600 can be hardware module or software mould
Block.For example, in certain embodiments, device 600 can with it is some or all of using software and
/ or firmware realize, such as be implemented as comprising computer program on a computer-readable medium
Product.Alternatively or additionally, device 600 can be realized with some or all of based on hardware,
Such as be implemented as integrated circuit (IC), application specific integrated circuit (ASIC), on-chip system (SOC),
Field programmable gate array (FPGA) etc..The scope of the present disclosure is not limited in this respect.
Fig. 7 is shown in which that showing for one or more realizations of subject matter described herein can be implemented
The block diagram of example computing system/server 700.Model training systems 110, model application system 120
Or both can be realized by computing system/server 700.Computing system/clothes shown in Fig. 7
Business device 700 is only example, and it should not form the function of using to realization described herein
With the limitation of scope.
As shown in fig. 7, computing system/server 700 is the form of universal computing device.Calculate
The component of systems/servers 700 can include but is not limited to one or more processors or processing
Unit 700, memory 720, one or more input equipments 730, one or more output are set
Standby 740, storage device 750, and one or more communication units 760.Processing unit 700
It can be reality or virtual processor and can persistently be held according to what is stored in memory 720
The various processing of row.In multiprocessing system, multiplied unit performs computer executable instructions,
To increase disposal ability.
Computing system/server 700 generally includes multiple computer medias.Such medium can be with
It is the addressable any medium that can be obtained of computing system/server 700, includes but is not limited to
Volatibility and non-volatile media, detachable and non-dismountable medium.Memory 720 can be
Volatile memory (such as register, cache, random access storage device (RAM)),
Non-volatile memory is (for example, read-only storage (ROM), electrically erasable are read-only
Memory (EEPROM), flash memory) or their certain combination.Storage device 750 can be with
It is detachable or non-dismountable, and machine readable media can be included, such as flash drive,
Disk or any other medium, it can be used in storage information and can be calculating
It is accessed in system/server 700.
Computing system/server 700 may further include it is other it is detachable/non-dismountable,
Volatile/non-volatile computer system storage medium.Although not shown in FIG. 7, Ke Yiti
For for from detachable, non-volatile magnetic disk (such as " floppy disk ") being read out or writing
Disk drive and the disc drives for being read out or writing from detachable, anonvolatile optical disk.
In such cases, each driving can be connected to always by one or more data media interfaces
Line.Memory 720 can include at least one program product, have (for example, at least one)
Program module collection, these program modules are configured as performing various realizations described herein
Function.
Program/utility program instrument of set with (for example, at least one) program module 724
722 can be stored in such as memory 720.Such program module 724 can include
But it is not limited to operating system, one or more application programs, other program modules and operation data.
Each example or particular combination in these examples can include the realization of networked environment.Program mould
Block 724 generally performs the function and/or method of the realization of subject matter described herein, such as method
200th, 300,400 and/or method 500.
Input block 730 can be one or more various input equipments.For example, input block
739 can include user equipment, mouse, keyboard, trackball etc..Communication unit 760
Realization is communicated to other computational entity on the communication media.Additionally, computing system/
The function of the component of server 600 can realize with single computing cluster or multiple computing machines,
These computing machines can fetch communication by communication link.Therefore, computing system/server 700
It can use and other one or more servers, NetPC Network PC (PC) or another
The logic of individual general networking node is connected to be operated in networked environment.Such as, but not limited to,
Communication media includes wired or Wireless networking technologies.
Computing system/server 700 can also as needed with one or more external equipments (not
Show) communicated, external equipment such as storage device, display device etc., with one or
It is multiple make it that user is communicated with the equipment that computing system/server 700 interacts, or with making
Any equipment of systems/servers 700 and other one or more computing device communications must be calculated
(for example, network interface card, modem etc.) is communicated.Such communication can be via input
/ (I/O) interface (not shown) is exported to perform.
Pass through above description, it will be understood that the realization of subject matter described herein need not be to noise
Statistical nature is made any it is assumed that but from scanned noise data and its corresponding true number
Truly counted according to how noise data is mapped to corresponding to it by the training data learning formed
According to.Therefore, the realization of subject matter described herein can be directed to any 3D scanners and can
Easily remove unknown noise.Proposed in the realization of subject matter described herein be referred to as through
The characteristic vector of the normal direction descriptor of filtering is robust for various noises.Described herein
Determine noise normal direction to very by training the regression tree of cascade proposed in the realization of theme
The mapping process of real normal direction is quick, and caused regression model can effectively remove divisor
Noise in.
Function described herein can be at least in part by one or more hardware logic groups
Part performs.Such as, but not limited to, the exemplary types bag for the hardware logic component that can be used
Include field programmable gate array (FPGA), application specific integrated circuit (ASIC), application specific standard
Product (ASSP), on-chip system (SOC), CPLD (CPLD)
Deng.
The program code of method for implementing subject matter described herein can use one or more
Any combinations of individual programming language are write.These program codes can be supplied to all-purpose computer,
The processor or controller of special-purpose computer or other programmable data processing units so that program
Code when by make when processor or controller execution the function of defined in flow chart and/or block diagram/
Operation is carried out.Program code can be performed completely on machine, partly performed on machine,
As independent software package partly on machine perform and partly on the remote machine perform or
Performed completely on remote machine or server.
In the context of present disclosure, machine readable media can be tangible medium, its
It can include or be stored for instruction execution system, device or equipment and perform system using or with instruction
The program that system, device or equipment are used in combination with.Machine readable media can be machine readable letter
Number medium or machine-readable storage medium.Machine readable media can include but is not limited to electronics,
Magnetic, optical, electromagnetism, infrared or semiconductor system, device or equipment, or
Any appropriate combination of person's the above.The more specific example of machinable medium can include
Electrical connection, portable computer diskette, hard disk, arbitrary access based on one or more lines are deposited
Reservoir (RAM), read-only storage (ROM), Erasable Programmable Read Only Memory EPROM (EPROM
Or flash memory), optical fiber, portable compact disk read-only storage (CD-ROM), light
Learn any appropriate combination of storage facilities, magnetic storage facilities or the above.
Although in addition, depicting each operation using certain order, this should be understood to requirement
So operation performs with shown certain order or in sequential order, or requires all diagrams
Operation should be performed to obtain desired result.Under certain environment, multitask and parallel place
Reason is probably favourable.Similarly, it is thin although containing some specific implementations in being discussed above
Section, but these are not construed as the limitation to the scope of subject matter described herein.In list
Some features described in the context of only realization can also be realized in single realization in combination
In.On the contrary, various features described in the context of single realization can also individually or
Realize in any suitable subcombination in multiple realizations.
It is listed below some example implementations of subject matter described herein.
In the first aspect, there is provided a kind of electronic equipment.The equipment includes at least one processing
Unit and at least one memory.At least one memory is coupled at least one processing unit
And store the instruction for being performed by least one processing unit.The instruction is when by least one
When individual processing unit performs so that the equipment:Obtain the noise normal direction of the first dough sheet in grid
With true normal direction;It is true from multiple neighbouring dough sheets of the first dough sheet in grid based on noise normal direction
First geometric properties of fixed first dough sheet;And based on the first geometric properties and true normal direction, really
The mapping of the true normal direction of fixed first geometric properties to the first dough sheet, for grid noise reduction.
In some implementations, determine that the first geometric properties include:Using wave filter to the first dough sheet
Normal direction change and change in location between multiple neighbouring dough sheets are filtered, and the is represented to generate
The characteristic vector of one geometric properties.
In some implementations, wave filter includes two-sided filter.
In some implementations, wave filter includes wave filter.
In some implementations, it is determined that mapping includes:Based on the first geometric properties and true normal direction,
Determine to map using regression analysis.
In some implementations, determine that mapping includes using regression analysis:It is special based on the first geometry
Seek peace true normal direction, generation represents the first regression tree of mapping, and the first regression tree is by the first geometry
First normal direction of the Feature Mapping to the true normal direction of expression.
In some implementations, generating the first regression tree includes being iteratively performed following operation at least one
It is secondary, counted out until the leaf segment of the first regression tree and reach predetermined threshold:In the every of the first regression tree
At individual node, the first regression model is trained using the training dataset associated with the node,
Training dataset is including the first geometric properties of each dough sheet in multiple dough sheets in grid and very
Real normal direction;Multiple training data subsets are obtained from training dataset;Utilize multiple training datas
Collect to train multiple second regression models;And missed in response to the fitting of multiple second regression models
Poor sum is less than predictive error threshold value, generates multiple child nodes of the node.
In some implementations, determine that mapping also includes using regression analysis:Based on the first normal direction,
Determine the second geometric properties of the first dough sheet;And based on the second geometric properties and true normal direction,
Second regression tree of generation and the cascade of the first regression tree, the second regression tree reflect the second geometric properties
The second normal direction for representing true normal direction is mapped to, the error between the second normal direction and true normal direction is less than
Error between first normal direction and true normal direction.
In some implementations, when being performed by least one processing unit, also causing should for the instruction
Equipment:Obtain the noise normal direction of the second dough sheet in grid;Noise normal direction based on the second dough sheet,
The 3rd geometric properties of the second dough sheet are determined from multiple neighbouring dough sheets of the second dough sheet in grid;
And based on the 3rd geometric properties and mapping, determine the true normal direction of the second dough sheet.
In second aspect, there is provided a kind of method to the grid noise reduction including dough sheet.The party
Method includes:Obtain the noise normal direction of the first dough sheet in grid and true normal direction;Based on Noise Method
To determining that the first geometry of the first dough sheet is special from multiple neighbouring dough sheets of the first dough sheet in grid
Sign;And based on the first geometric properties and true normal direction, determine the first geometric properties to the first face
The mapping of the true normal direction of piece, for grid noise reduction.
In some implementations, determine that the first geometric properties include:To the first dough sheet with it is multiple neighbouring
Normal direction change and change in location between dough sheet are filtered, and the first geometric properties are represented with generation
Characteristic vector.
In some implementations, filtering includes:To the method between the first dough sheet and multiple neighbouring dough sheets
Bilateral filtering is carried out to change and change in location.
In some implementations, filtering includes:To the method between the first dough sheet and multiple neighbouring dough sheets
Filtering is guided to change and change in location.
In some implementations, it is determined that mapping includes:Based on the first geometric properties and true normal direction,
Determine to map using regression analysis.
In some implementations, determine that mapping includes using regression analysis:It is special based on the first geometry
Seek peace true normal direction, generation represents the first regression tree of mapping, and the first regression tree is by the first geometry
First normal direction of the Feature Mapping to the true normal direction of expression.
In some implementations, generating the first regression tree includes being iteratively performed following operation at least one
It is secondary, counted out until the leaf segment of the first regression tree and reach predetermined threshold:In the every of the first regression tree
At individual node, the first regression model is trained using the training dataset associated with the node,
Training dataset is including the first geometric properties of each dough sheet in multiple dough sheets in grid and very
Real normal direction;Multiple training data subsets are obtained from training dataset;Utilize multiple training datas
Collect to train multiple second regression models;And missed in response to the fitting of multiple second regression models
Poor sum is less than predictive error threshold value, generates multiple child nodes of the node.
In some implementations, determine that mapping also includes using regression analysis:Based on the first normal direction,
Determine the second geometric properties of the first dough sheet;And based on the second geometric properties and true normal direction,
Second regression tree of generation and the cascade of the first regression tree, the second regression tree reflect the second geometric properties
The second normal direction for representing true normal direction is mapped to, the error between the second normal direction and true normal direction is less than
Error between first normal direction and true normal direction.
In some implementations, this method also includes:Obtain the Noise Method of the second dough sheet in grid
To;Noise normal direction based on the second dough sheet, from multiple neighbouring dough sheets of the second dough sheet in grid
Determine the 3rd geometric properties of the second dough sheet;And based on the 3rd geometric properties and mapping, it is determined that
The true normal direction of second dough sheet.
In a third aspect, there is provided a kind of computer-readable recording medium.This is computer-readable
Storage medium has the computer-readable program instructions being stored thereon.Computer-readable program refers to
The step of making for performing the method according to first aspect.
Although described using the language specific to architectural feature and/or method logical action
This theme it should be appreciated that the theme defined in appended claims is not necessarily limited to
The special characteristic of face description or action.On the contrary, special characteristic described above and action are only
It is the exemplary forms for realizing claims.
Claims (20)
1. a kind of electronic equipment, including:
At least one processing unit;
At least one memory, at least one memory are coupled at least one place
Manage unit and store the instruction for being performed by least one processing unit, the instruction
When being performed by least one processing unit so that the equipment:
Obtain the noise normal direction of the first dough sheet in grid and true normal direction;
Based on the noise normal direction, from multiple neighbours of first dough sheet in the grid
Nearly dough sheet determines the first geometric properties of first dough sheet;And
Based on first geometric properties and the true normal direction, first geometry is determined
Feature to the true normal direction of first dough sheet mapping, for the grid noise reduction.
2. equipment according to claim 1, wherein determining that first geometric properties include:
The normal direction between first dough sheet and the multiple neighbouring dough sheet is become using wave filter
Change and change in location is filtered, the characteristic vector of first geometric properties is represented with generation.
3. equipment according to claim 2, wherein the wave filter includes two-sided filter.
4. equipment according to claim 2, wherein the wave filter includes wave filter.
5. equipment according to claim 1, wherein determining that the mapping includes:
Based on first geometric properties and the true normal direction, using regression analysis to determine
State mapping.
6. equipment according to claim 5, wherein being reflected described in being determined using regression analysis
Penetrate including:
Based on first geometric properties and the true normal direction, generation represents the of the mapping
One regression tree, first geometric properties are mapped to by first regression tree represents described true
First normal direction of normal direction.
7. equipment according to claim 6, wherein generating first regression tree includes:
Following operation is iteratively performed at least once, until the leaf segment of first regression tree is counted
Mesh reaches predetermined threshold:
At each node of first regression tree, using associated with the node
Training dataset trains the first regression model, and the training dataset includes in the grid
First geometric properties of each dough sheet and the true normal direction in multiple dough sheets;
Multiple training data subsets are obtained from the training dataset;
Multiple second regression models are trained using the multiple training data subset;And
It is less than predetermined miss in response to the error of fitting sum of the multiple second regression model
Poor threshold value, generate multiple child nodes of the node.
8. equipment according to claim 5, wherein being reflected described in being determined using regression analysis
Penetrating also includes:
Based on first normal direction, the second geometric properties of first dough sheet are determined;And
Based on second geometric properties and the true normal direction, generation and first regression tree
Second geometric properties are mapped to expression by the second regression tree of cascade, second regression tree
Second normal direction of the true normal direction, the error between second normal direction and the true normal direction
Less than the error between first normal direction and the true normal direction.
9. equipment according to claim 1, wherein the instruction is when by described at least one
When processing unit performs, also cause the equipment:
Obtain the noise normal direction of the second dough sheet in the grid;
The noise normal direction based on second dough sheet, from second face in the grid
Multiple neighbouring dough sheets of piece determine the 3rd geometric properties of second dough sheet;And
Based on the 3rd geometric properties and the mapping, the true method of second dough sheet is determined
To.
10. a kind of method to the grid noise reduction including dough sheet, including:
Obtain the noise normal direction of the first dough sheet in grid and true normal direction;
Based on the noise normal direction, from multiple vicinal faces of first dough sheet in the grid
Piece determines the first geometric properties of first dough sheet;And
Based on first geometric properties and the true normal direction, first geometric properties are determined
To the mapping of the true normal direction of first dough sheet, for the grid noise reduction.
11. according to the method for claim 10, wherein determining the first geometric properties bag
Include:
Normal direction change between first dough sheet and the multiple neighbouring dough sheet and position are become
Change is filtered, and the characteristic vector of first geometric properties is represented with generation.
12. according to the method for claim 11, wherein the filtering includes:
Normal direction change between first dough sheet and the multiple neighbouring dough sheet and position are become
Change and carry out bilateral filtering.
13. according to the method for claim 11, wherein the filtering includes:
Normal direction change between first dough sheet and the multiple neighbouring dough sheet and position are become
Change guides filtering.
14. according to the method for claim 10, wherein determining that the mapping includes:
Based on first geometric properties and the true normal direction, using regression analysis to determine
State mapping.
15. according to the method for claim 14, wherein described in being determined using regression analysis
Mapping includes:
Based on first geometric properties and the true normal direction, generation represents the of the mapping
One regression tree, first geometric properties are mapped to by first regression tree represents described true
First normal direction of normal direction.
16. according to the method for claim 15, wherein generating first regression tree includes:
Following operation is iteratively performed at least once, until the leaf segment of first regression tree is counted
Mesh reaches predetermined threshold:
At each node of first regression tree, using associated with the node
Training dataset trains the first regression model, and the training dataset includes in the grid
First geometric properties of each dough sheet and the true normal direction in multiple dough sheets;
Multiple training data subsets are obtained from the training dataset;
Multiple second regression models are trained using the multiple training data subset;And
It is less than predetermined miss in response to the error of fitting sum of the multiple second regression model
Poor threshold value, generate multiple child nodes of the node.
17. according to the method for claim 15, wherein described in being determined using regression analysis
Mapping also includes:
Based on first normal direction, the second geometric properties of first dough sheet are determined;And
Based on second geometric properties and the true normal direction, generation and first regression tree
Second geometric properties are mapped to expression by the second regression tree of cascade, second regression tree
Second normal direction of the true normal direction, the error between second normal direction and second normal direction
Less than the error between first normal direction and second normal direction.
18. the method according to claim 11, in addition to:
Obtain the noise normal direction of the second dough sheet in the grid;
The noise normal direction based on second dough sheet, from second face in the grid
Multiple neighbouring dough sheets of piece determine the 3rd geometric properties of second dough sheet;And
Based on the 3rd geometric properties and the mapping, the true method of second dough sheet is determined
To.
19. a kind of computer-readable recording medium, the computer-readable recording medium, which has, to be deposited
In computer-readable program instructions thereon, the computer-readable program instructions are used to perform for storage:
Obtain the noise normal direction of the first dough sheet in grid and true normal direction;
Based on the noise normal direction, from multiple vicinal faces of first dough sheet in the grid
Piece determines the first geometric properties of first dough sheet;And
Based on first geometric properties and the true normal direction, first geometric properties are determined
To the mapping of the true normal direction of first dough sheet, for the grid noise reduction.
20. computer-readable recording medium according to claim 19, wherein the calculating
Machine readable program instructions are additionally operable to perform:
Obtain the noise normal direction of the second dough sheet in the grid;
The noise normal direction based on second dough sheet, from second face in the grid
Multiple neighbouring dough sheets of piece determine the 3rd geometric properties of second dough sheet;And
Based on the 3rd geometric properties and the mapping, the true method of second dough sheet is determined
To.
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610285481.4A CN107341768B (en) | 2016-04-29 | 2016-04-29 | Grid noise reduction |
PCT/CN2017/078352 WO2017185937A1 (en) | 2016-04-29 | 2017-03-28 | Mesh denoising |
US16/096,448 US10762657B2 (en) | 2016-04-29 | 2017-03-28 | Mesh denoising |
EP17788590.2A EP3449462B1 (en) | 2016-04-29 | 2017-03-28 | Mesh denoising |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610285481.4A CN107341768B (en) | 2016-04-29 | 2016-04-29 | Grid noise reduction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107341768A true CN107341768A (en) | 2017-11-10 |
CN107341768B CN107341768B (en) | 2022-03-11 |
Family
ID=60161849
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610285481.4A Active CN107341768B (en) | 2016-04-29 | 2016-04-29 | Grid noise reduction |
Country Status (4)
Country | Link |
---|---|
US (1) | US10762657B2 (en) |
EP (1) | EP3449462B1 (en) |
CN (1) | CN107341768B (en) |
WO (1) | WO2017185937A1 (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110930334A (en) * | 2019-11-26 | 2020-03-27 | 浙江大学 | Grid denoising method based on neural network |
CN113178013A (en) * | 2021-05-24 | 2021-07-27 | 广东奥普特科技股份有限公司 | Triangular mesh filtering method and device, electronic equipment and storage medium |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111028356A (en) * | 2019-11-25 | 2020-04-17 | 中国地质大学(武汉) | Optimization method based on non-convex non-smooth second-order regular term and sparse fidelity term |
CN112529803B (en) * | 2020-12-10 | 2022-05-13 | 深圳市数字城市工程研究中心 | Feature-preserving three-dimensional Mesh model denoising method |
CN114419275A (en) * | 2021-12-13 | 2022-04-29 | 深圳先进技术研究院 | Method for denoising triangular mesh based on dual-graph neural network |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101877148A (en) * | 2009-11-13 | 2010-11-03 | 北京航空航天大学 | Method for repairing three-dimensional grid model based on global structure |
CN102880700A (en) * | 2012-09-24 | 2013-01-16 | 常州大学 | Three dimensional point model data denoising method |
CN103021017A (en) * | 2012-12-04 | 2013-04-03 | 上海交通大学 | Three-dimensional scene rebuilding method based on GPU acceleration |
US20130321593A1 (en) * | 2012-05-31 | 2013-12-05 | Microsoft Corporation | View frustum culling for free viewpoint video (fvv) |
WO2014190037A1 (en) * | 2013-05-22 | 2014-11-27 | Siemens Aktiengesellschaft | Parameter estimation for mesh segmentation using random walks |
CN104282038A (en) * | 2014-09-26 | 2015-01-14 | 北京航空航天大学 | Point-cloud feature point extraction method based on local sub-neighborhood division |
CN104282040A (en) * | 2014-09-29 | 2015-01-14 | 北京航空航天大学 | Finite element preprocessing method for reconstructing three-dimensional entity model |
CN105096268A (en) * | 2015-07-13 | 2015-11-25 | 西北农林科技大学 | Denoising smoothing method of point cloud |
Family Cites Families (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6987511B2 (en) * | 2002-10-17 | 2006-01-17 | International Business Machines Corporation | Linear anisotrophic mesh filtering |
US7519488B2 (en) | 2004-05-28 | 2009-04-14 | Lawrence Livermore National Security, Llc | Signal processing method and system for noise removal and signal extraction |
US7589720B2 (en) * | 2004-08-04 | 2009-09-15 | Microsoft Corporation | Mesh editing with gradient field manipulation and user interactive tools for object merging |
US8698800B2 (en) * | 2005-11-29 | 2014-04-15 | Siemens Corporation | Method and apparatus for non-shrinking mesh smoothing using local fitting |
US9792728B2 (en) * | 2012-08-17 | 2017-10-17 | Thomson Licensing | Method and apparatus for 3D mesh de-noising |
US8908989B2 (en) * | 2012-11-02 | 2014-12-09 | Board Of Regents, The University Of Texas System | Recursive conditional means image denoising |
CN104574282B (en) * | 2013-10-22 | 2019-06-07 | 鸿富锦精密工业(深圳)有限公司 | Point cloud noise spot removes system and method |
US9344690B2 (en) | 2014-01-24 | 2016-05-17 | Microsoft Technology Licensing, Llc | Image demosaicing |
CN106133788B (en) * | 2014-11-13 | 2019-11-12 | 华为技术有限公司 | Handle the image processing apparatus of digital picture |
CN104952107A (en) | 2015-05-18 | 2015-09-30 | 湖南桥康智能科技有限公司 | Three-dimensional bridge reconstruction method based on vehicle-mounted LiDAR point cloud data |
US20170178390A1 (en) * | 2015-12-18 | 2017-06-22 | Canon Kabushiki Kaisha | Devices, systems, and methods for measuring and reconstructing the shapes of specular objects by multiview capture |
US10559085B2 (en) * | 2016-12-12 | 2020-02-11 | Canon Kabushiki Kaisha | Devices, systems, and methods for reconstructing the three-dimensional shapes of objects |
-
2016
- 2016-04-29 CN CN201610285481.4A patent/CN107341768B/en active Active
-
2017
- 2017-03-28 WO PCT/CN2017/078352 patent/WO2017185937A1/en active Application Filing
- 2017-03-28 US US16/096,448 patent/US10762657B2/en active Active
- 2017-03-28 EP EP17788590.2A patent/EP3449462B1/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101877148A (en) * | 2009-11-13 | 2010-11-03 | 北京航空航天大学 | Method for repairing three-dimensional grid model based on global structure |
US20130321593A1 (en) * | 2012-05-31 | 2013-12-05 | Microsoft Corporation | View frustum culling for free viewpoint video (fvv) |
CN102880700A (en) * | 2012-09-24 | 2013-01-16 | 常州大学 | Three dimensional point model data denoising method |
CN103021017A (en) * | 2012-12-04 | 2013-04-03 | 上海交通大学 | Three-dimensional scene rebuilding method based on GPU acceleration |
WO2014190037A1 (en) * | 2013-05-22 | 2014-11-27 | Siemens Aktiengesellschaft | Parameter estimation for mesh segmentation using random walks |
CN104282038A (en) * | 2014-09-26 | 2015-01-14 | 北京航空航天大学 | Point-cloud feature point extraction method based on local sub-neighborhood division |
CN104282040A (en) * | 2014-09-29 | 2015-01-14 | 北京航空航天大学 | Finite element preprocessing method for reconstructing three-dimensional entity model |
CN105096268A (en) * | 2015-07-13 | 2015-11-25 | 西北农林科技大学 | Denoising smoothing method of point cloud |
Non-Patent Citations (5)
Title |
---|
JAMES R. DIEBEL AND SEBASTIAN THRUN: "A Bayesian Method for Probable Surface Reconstruction and Decimation", 《ACM TRANSACTIONS ON GRAPHICS,》 * |
JINZE YU ET AL.: "Feature-preserving mesh denoising via normal guided quadric error metrics", 《OPTICS AND LASERS IN ENGINEERING》 * |
JUN WANG ET AL.: "A cascaded approach for feature-preserving surface mesh denoising", 《COMPUTER-AIDED DESIGN》 * |
周浩宇: "基于生物视觉模型的图形图像处理方法", 《中国博士学位论文全文数据库 信息科技辑》 * |
罗立彦: "图像插值与离散曲面去噪", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110930334A (en) * | 2019-11-26 | 2020-03-27 | 浙江大学 | Grid denoising method based on neural network |
CN113178013A (en) * | 2021-05-24 | 2021-07-27 | 广东奥普特科技股份有限公司 | Triangular mesh filtering method and device, electronic equipment and storage medium |
CN113178013B (en) * | 2021-05-24 | 2023-10-03 | 广东奥普特科技股份有限公司 | Triangular mesh filtering method, triangular mesh filtering device, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
WO2017185937A1 (en) | 2017-11-02 |
EP3449462A1 (en) | 2019-03-06 |
US20190139248A1 (en) | 2019-05-09 |
CN107341768B (en) | 2022-03-11 |
US10762657B2 (en) | 2020-09-01 |
EP3449462A4 (en) | 2019-12-04 |
EP3449462B1 (en) | 2021-03-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Shiri et al. | A comprehensive overview and comparative analysis on deep learning models: CNN, RNN, LSTM, GRU | |
US11501192B2 (en) | Systems and methods for Bayesian optimization using non-linear mapping of input | |
Zhao et al. | Classifying airborne LiDAR point clouds via deep features learned by a multi-scale convolutional neural network | |
EP3686848A1 (en) | Semantic image synthesis for generating substantially photorealistic images using neural networks | |
CN107341768A (en) | Grid noise reduction | |
Xie et al. | Generative VoxelNet: Learning energy-based models for 3D shape synthesis and analysis | |
US20130177235A1 (en) | Evaluation of Three-Dimensional Scenes Using Two-Dimensional Representations | |
CN108182260B (en) | Multivariate time sequence classification method based on semantic selection | |
CN110348399A (en) | EO-1 hyperion intelligent method for classifying based on prototype study mechanism and multidimensional residual error network | |
CN110909125B (en) | Detection method of media rumor of news-level society | |
JP2020522773A (en) | Detection and representation of objects in images | |
Pietikäinen et al. | Challenges of Artificial Intelligence--From Machine Learning and Computer Vision to Emotional Intelligence | |
US20220207354A1 (en) | Analog circuits for implementing brain emulation neural networks | |
CN115222896B (en) | Three-dimensional reconstruction method, three-dimensional reconstruction device, electronic equipment and computer readable storage medium | |
CN109242106A (en) | sample processing method, device, equipment and storage medium | |
Bazarghan | Application of self-organizing map to stellar spectral classifications | |
Salazar | On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling | |
Gupta et al. | 3d point cloud feature explanations using gradient-based methods | |
CN111026267A (en) | VR electroencephalogram idea control interface system | |
Sanghvi et al. | Fauna image classification using convolutional neural network | |
Xu et al. | A novel image feature extraction algorithm based on the fusion AutoEncoder and CNN | |
Taslim et al. | Plant leaf identification system using convolutional neural network | |
CN108073934A (en) | Nearly multiimage detection method and device | |
Fortuna-Cervantes et al. | Evaluation of CNN Models with Transfer Learning in Art Media Classification in Terms of Accuracy and Class Relationship | |
CN109086651A (en) | Handwriting model training method, hand-written character recognizing method, device, equipment and medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |